Clinical Utility Assessment for Physics-Informed Seizure Prediction

Assessment results evaluate a Hybrid Physics-Informed Graph Neural Network for predicting epileptic seizures using EEG data. The model was tested on 22 pediatric patients from the CHB-MIT dataset and 16 adults from IEEG.org, achieving 84.7% chimera detection accuracy. Author Masoud Amiri published the findings in April 2026.

Use Cases

• Compare chimera detection accuracy (84.7%) against baseline methods like Delay Differential Analysis (75.5%).
• Analyze cross-patient generalization performance (79.8%) for model robustness evaluation.
• Examine the relationship between physics-informed regularization (weight λ₁=0.03) and training data efficiency (35% reduction).
• Validate seizure prediction sensitivity (89.3%) and false positive rate (0.48 per hour) at 90-minute horizons.
• Assess zero-shot transfer accuracy (71.3%) from scalp to intracranial EEG recordings.

Strengths

• Results are derived from evaluation on two established EEG datasets: CHB-MIT (22 patients, 182 seizures) and IEEG.org (16 patients, 87 seizures).
• Key performance metrics are provided with confidence intervals, such as 84.7% chimera detection accuracy (95% CI: 82.3–87.1%).
• The assessment includes multiple validation aspects: cross-patient generalization, data efficiency, and zero-shot transfer.

Limitations

• The dataset is very small at 5.5 KB, containing only summary assessment results, not raw EEG data or model outputs.
• The clinical utility results are specific to the proposed HP-GNN architecture and may not generalize to other prediction methods.
• Temporal coverage and specific patient demographics beyond 'pediatric' and 'adult' are not detailed in the available information.

Provenance

Source

figshare, authored by Masoud Amiri.

Collection Method

Results from evaluating a novel Hybrid Physics-Informed Graph Neural Network (HP-GNN) on EEG data.

Freshness

Last updated April 2026.